Developing and deploying a use-inspired metapopulation framework for stratified health outcomes

People involved

Roadmap

• Motivation & collaboration context
  
• MetaRVM modeling framework
  • Metapopulation structure
    
  • Disease progression model
    
  • Mixing matrices from a synthetic population
    
• Implementation as an R package
• Trajectory-oriented optimization
• Case study: influenza in Chicago

Motivation

• Public health decision makers increasingly rely on epidemiological models to:
  • Forecast likely futures
  • Stress-test interventions
  • Plan resource allocation (e.g., beds, staffing)

Motivation

Motivation

MetaRVM

• An open-source R package for modeling infectious disease spread across stratified subpopulations.
• Subpopulations can be defined by:
  • Geography (e.g., neighborhood, ZIP, zone)
  • Demographics (e.g., age, race/ethnicity)
• Features:
  • Time varying mixing pattern
  • Extended SEIR-type disease progression
  • Checkpointing functionality
  • Integration with Trajectory-Oriented Optimization workflows

Model structure: extended SEIR

Core disease states (example):

• S – Susceptible
  
• E – Exposed (infected, not yet infectious)
  
• I – Infectious (possibly with severity subcompartments)
  
• H – Hospitalized
  
• R – Recovered / immune
  
• V – Vaccinated

Key design choices:

• Discrete-time simulation, typically daily time steps
• Transitions governed by:
  • Latent and infectious periods
  • Age- and stratum-specific risks
  • Vaccination efficacy and waning, if modeled
• Outputs:
  • Incident infections, hospitalizations, etc.
  • Cumulative outcomes by stratum

Force of infection and mixing

• For each stratum, the force of infection depends on:
  • Contact rates between strata
  • Prevalence of infectious individuals in each stratum
• Let:
  • β(t) be a time-varying transmission parameter
  • M(t) be a mixing matrix (strata × strata) at time t
  • I(t) the vector of infectious counts by stratum
• Conceptually:
  • Infection pressure in stratum k combines:
    • Its own infectious individuals
    • Contacts with infectious individuals in other strata
  • With separate terms for susceptible vs. vaccinated if modeled

Implementation uses efficiently stored and processed mixing matrices derived from a synthetic population.

Building mixing matrices from a synthetic population

• Underlying asset: synthetic population for Chicago
  • Statistically representative of residents and households
  • Includes:
    • Demographics (age, etc.)
    • Home, work, school, and activity locations
    • Daily schedules / contact opportunities
• From this, we construct empirical mixing matrices:
  • Count cross-stratum contacts in:
    • Households
    • Schools
    • Workplaces
    • Other locations
  • Aggregate to daily or weekly matrices
  • Separate weekday / weekend patterns

This leverages the richness of CityCOVID’s synthetic ecosystem, but in a metapopulation abstraction.

Example: weekday vs weekend mixing

• Weekday:
  • Strong school-based mixing among children
  • Workplace mixing among working-age adults
  • Cross-age mixing in households in mornings and evenings
• Weekend:
  • Less structured mixing
  • More household and community mixing
• MetaRVM:
  • Can use separate mixing matrices for:
    • Weekdays vs weekends
    • Different phases of the season
    • Intervention periods (e.g., school closures)

Workflow: from data to decision

1. Define stratification
   • e.g., age × neighborhood, or age × race
2. Prepare inputs
   • Population counts by stratum
   • Daily or weekly mixing matrices
   • Vaccination and other intervention schedules
3. Configure MetaRVM
   • Structural and epidemiological parameters
   • Simulation horizon and time step
4. Calibrate to observed data
   • Hospitalizations, case counts, etc.
   • Use Bayesian optimization / calibration tools
5. Run scenarios
   • Alternative vaccination strategies
   • Timing and intensity of NPIs
6. Summarize and communicate
   • Visualizations
   • Key metrics for decision makers

Case study: influenza hospitalizations in Chicago (2023–24)

• Objective:
  • Support prospective surveillance of influenza-related hospitalizations in Chicago
• Setup:
  • Strata: age groups (e.g., children, adults, seniors)
  • Observed data:
    • Daily or weekly influenza-related hospitalizations
  • Inputs:
    • Synthetic population-based mixing
    • Empirical vaccination curves over the season
• Tasks:
  • Calibrate MetaRVM to observed hospitalizations
  • Generate short-term forecasts
  • Explore alternative vaccination scenarios

Calibration: overview

• Calibration goal:
  • Align model-generated hospitalizations with observed time series
• Approach:
  • Choose a set of calibration parameters (e.g., transmission scaling, reporting/observation fraction)
  • Define a loss function (e.g., squared error between modeled and observed hospitalizations)
  • Use Bayesian optimization (or other search strategies) to:
    • Explore parameter space efficiently
    • Balance exploitation and exploration
• Outcome:
  • Posterior-like distribution over plausible parameter values
  • Uncertainty envelopes around model trajectories

Limitations and extensions

Limitations:

• Still an aggregated model:
  • Does not capture individual-level heterogeneity beyond stratification
• Dependent on:
  • Quality of the synthetic population and mixing matrices
  • Quality and timeliness of surveillance data
• Calibration challenges:
  • Parameter identifiability
  • Structural uncertainty

Possible extensions:

• Co-circulation of multiple pathogens (e.g., flu + RSV + COVID-19)
• More nuanced interventions:
  • Layered NPIs, dynamic behavior change
  • Targeted vaccination by geography and risk
• Deeper integration with:
  • High-performance computing resources
  • Automated data ingestion pipelines

Summary

• MetaRVM provides a use-inspired metapopulation modeling framework for:
  • Detailed, stratified tracking of health outcomes
  • Infectious disease spread across multiple interacting subpopulations
• Key contributions:
  • Uses synthetic population-derived mixing matrices
  • Balances realism with computational efficiency
  • Delivered as an open-source R package with a Shiny front end
• Case study:
  • Demonstrated on influenza-related hospitalizations in Chicago
  • Supports prospective surveillance and scenario analysis

Looking ahead

• Research directions:
  • Multi-pathogen modeling
  • Enhanced calibration and uncertainty quantification
  • Integration of behavioral and policy feedbacks
• Deployment directions:
  • Broader adoption at other health departments
  • Sharing of configuration templates and best practices
  • Training materials for non-modeling stakeholders

Acknowledgments

• Chicago Department of Public Health
  
• Collaborators at Argonne National Laboratory and University of Chicago
  
• Funding and institutional support as detailed in the paper

Questions?

Thank you!